Graph Convolutional Network Research Articles

Alzheimer's disease classification using mutual information generated graph convolutional network for functional MRI.

BackgroundHigh-order cognitive functions depend on collaborative actions and information exchange between multiple brain regions. These inter-regional interactions can be characterized by mutual information (MI). Alzheimer's disease (AD) is known to affect many high-order cognitive functions, suggesting an alteration to inter-regional MI, which remains unstudied.ObjectiveTo examine whether inter-regional MI can effectively distinguish different stages of AD from normal control (NC) through a connectome-based graph convolutional network (GCN).MethodsMI was calculated between the mean time series of each pair of brain regions, forming the connectome which was input to a multi-level connectome based GCN (MLC-GCN) to predict the different stages of AD and NC. The spatio-temporal feature extraction in MLC-GCN was used to capture multi-level functional connectivity patterns generating connectomes. The GCN predictor learns and optimizes graph representations at each level, concatenating the representations for final classification. We validated our model on 552 subjects from ADNI and OASIS3. The MI-based model was compared to models with several different connectomes defined by Kullback-Leibler divergence, cross-entropy, cross-sample entropy, and correlation coefficient. Model performance was evaluated using 5-fold cross-validation.ResultsThe MI-based connectome achieved the highest prediction performance for both ADNI2 and OASIS3 where it's accuracy/Area Under the Curve/F1 were 87.72%/0.96/0.88 and 84.11%/0.96/0.91 respectively. Model visualization revealed that prominent MI features located in temporal, prefrontal, and parietal cortices.ConclusionsMI-based connectomes can reliably differentiate NC, mild cognitive impairment and AD. Compared to other four measures, MI demonstrated the best performance. The model should be further tested with other independent datasets.

Machine Learning-Driven Prediction of Corrosion Inhibitor Efficiency: Emerging Algorithms, Challenges, and Future Outlooks

Abstract Machine learning (ML) frameworks are transforming the development of corrosion inhibitors by enabling quantitative prediction of inhibition efficiency before synthesis. This work identifies the most reliable machine learning (ML) strategies for forecasting corrosion inhibitor efficiency before synthesis, thereby shortening development cycles and reducing experimental cost. Drawing on more than fifteen harmonized datasets that span pyrimidines, ionic liquids, graphene oxides, and additional compound families, we benchmark traditional algorithms, such as artificial neural networks, support vector machines, k-nearest neighbors, random forests, against advanced graph-based and deep architectures including three-level directed message-passing neural networks, 2D3DMol-CIC, and graph convolutional networks. Cohesive data collections exceeding four hundred molecules under standardized test conditions deliver coefficients of determination above 0.90 and root-mean-square errors below 0.05. In contrast, fragmented datasets suffer from overfitting with R2 often under 0.70. Graph neural networks lower prediction error by up to thirty percent relative to descriptor-driven QSAR models for structurally diverse inhibitors. However, their accuracy diminishes for large, flexible molecules unless explicit three-dimensional information is provided. Ensemble schemes such as Gaussian process regression with simple averaging and gradient boosting regressors fortified by permutation feature importance improve robustness in noisy or multi-alloy environments. At the same time, virtual sample augmentation and genetic algorithm feature selection elevate sparse data performance, raising k-nearest neighbor models from R2 = 0.05 to 0.99 in a representative thiophene set. Persistent obstacles include limited public databases, inconsistent experimental protocols, and the opaque nature of deep learners. Researchers, engineers, and material scientists will gain valuable insights into optimizing ML-driven corrosion predictions, guiding future experimental studies.

Multiple object tracking using weighted graph convolutional neural networks

Multiple object tracking using weighted graph convolutional neural networks

Knowledge Tracing Enhanced by Graph Convolutional Networks With Self-Supervised Learning

Knowledge tracing (KT) plays a key role in adaptive learning, yet traditional recurrent neural network–based methods often struggle with sparse data and overlook relationships between knowledge points. To address these limitations, this paper proposes a novel knowledge tracing model (KGS-KT) that integrates knowledge graphs, graph convolutional networks (GCNs), and self-supervised learning (SSL). Our approach constructs a knowledge graph from learners' performance data to capture inherent relationships among knowledge points and employs GCNs to generate comprehensive knowledge point embeddings. In addition, two SSL tasks—node attribute prediction and edge relation prediction—are introduced. These tasks enhance feature representation by reconstructing masked node attributes and inferring missing connections, while also providing deeper insight into the role of structured dependencies. This strengthens the model's robustness against data sparsity. Experimental results show that our approach outperforms existing methods, demonstrating the effectiveness of combining structured knowledge and SSL in KT.

Construction of a knowledge graph for dance teaching based on improved graph convolutional networks

Aiming at the inefficiency of retrieval and insufficient recommendation caused by the dispersion of dance multimodal resources, this study proposes a knowledge graph construction method based on improved GCNs. Integrating public dance libraries and teaching resources, the first dance knowledge graph DanceKG is constructed, covering entity relationships such as movements and styles. DRA-GCN is innovatively proposed to quantify the interaction frequency of nodes through dynamic weighting module and strengthen the characterization of long-tailed entities associated with complex movements by combining multi-head attention. Experiments show that DRA-GCN has a link prediction MRR of 0.72 (9% higher than traditional GCN) on 50,000 triad datasets, and the F1-value of movement classification is improved by 5%, which supports the movement recommendation and error correction of intelligent teaching system and promotes the digital development of dance education.

Deep Learning Framework for Oil Shale Pyrolysis State Recognition Using Bionic Electronic Nose

Real-time monitoring of the pyrolysis state of oil shale is crucial for precisely controlling heating temperature and duration, which can significantly reduce extraction costs. However, due to the complexity of in-situ environments, this task is highly challenging and remains one of the key technological barriers in in-situ mining. To address this issue, this paper proposes an end-to-end recognition technology solution for in-situ pyrolysis state of oil shale using electronic nose. The proposed solution integrates Graph Convolutional Network (GCN) and Long Short-Term Memory (LSTM) to capture the spatial correlations among different sensors in the electronic nose and the temporal characteristics of the data, respectively. It is designed to identify both the pyrolysis state classification and the oil shale maturity regression tasks. Our model achieves 93.87% accuracy on the task of classifying the pyrolysis stage of oil shale; the R2 on the regression task reaches 0.93. To evaluate its effectiveness, we compare its performance with state-of-the-art (SOTA) methods in this field. Experimental results demonstrate the superiority of our proposed framework, highlighting its effectiveness and advantages over existing methods.

Linear projection fused graph-based semi-supervised learning on multi-view data

In recent years, the surge in data-driven applications across various domains has spurred heightened interest in semi-supervised learning applied to graphs. This surge is attributed to the ubiquitous presence of graph data structures in real-world contexts, such as social networks’ interpersonal relationships, recommender systems’ user behavior graphs, and bioinformatics’ molecular interaction networks. However, for certain data types like images, not only is there a dearth of explicit graph structure, but also the existence of multiple view description methods complicates matters further. The intricacies of multi-view data pose challenges in directly applying traditional semi-supervised learning techniques to graphs. Consequently, researchers have begun exploring the fusion of semi-supervised learning with deep learning to leverage its wealth of information and enhance model efficacy. Effectively amalgamating graph structures with multi-view data remains a challenging problem necessitating further research. This paper introduces the Linear projection Fused Graph-based Semi-supervised Classification (LFGSC) method tailored for multi-view data, building upon the Graph Convolutional Network (GCN) architecture. Firstly, for each view, we leverage a semi-supervised approach that provides the concurrent estimation of the corresponding graph and the flexible linear data representations in a low-dimensional feature space. Subsequently, an adaptive and unified graph is generated, followed by the utilization of a fully connected network to fuse the projected features further and reduce dimensionality. Finally, the fused features and graph are inputted into a GCN to conduct semi-supervised classification. During training, the model incorporates cross-entropy loss, manifold regularization loss, graph auto-encoder loss, and supervised contrastive loss. Leveraging linear transformation significantly diminishes the input feature dimensions for GCN, thereby achieving high accuracy while substantially reducing computational overhead. Furthermore, experimental results conducted on various bench-marked multi-view image datasets demonstrate the superiority of LFGSC over existing semi-supervised learning methods for multi-view scenarios. (Source code: https://github.com/BiJingjun/LFGSC.)

Hybridisation of Graph Convolutional Neural Networks and Residual Attention Network for early identification of plant diseases

ABSTRACT Due to the limitation in the existing infrastructure, the recognition of plant disease has become a crucial task, because crop diseases are considered as the major threat in terms of food security in all parts of the world. Hence, earlier prevention and timely detection are regarded as the two major factors for effectively enhancing the rate of production. Hence, automated computational systems are designed for diagnosing plant leaf diseases. At first, the standard images of plant leaf are gathered from online sources and given into pre-processing phase. Next, the pre-processing is done using optimally weighted threshold histogram equalization, where weights are optimized using Enhanced Sandpiper Optimization Algorithm (ESOA). Outcomes are then, directly fed to the Hybrid deep learning strategy with the hybridization of Residual Attention network (RAN) and Graph Convolutional Neural Networks (GCNN) termed as GCNN-RAN, in which the same E-SOA is used for optimizing the parameters of GCNN and RAN. The accuracy and precision rate obtained from the recommended approach are 97.33% and 96.5%, which shows the disease detection in plant leaf accuracy is more impressive than previously developed models. Thus, the newly developed model is useful for detecting the disease over the plant leaf in an effective manner.

Hyperbolic Bernstein Neural Networks: Enhancing graph convolutions in non-Euclidean spaces.

Hyperbolic Bernstein Neural Networks: Enhancing graph convolutions in non-Euclidean spaces.

Group behavioral intention recognition based on semantic relations analysis

Group Activity Recognition (GAR) is the task of recognizing an overall activity in a multi-individual scene. Most of the existing methods have achieved significant progress by incorporating the attributes and relations between individuals. However, these methods still suffer from the ability to automatically detect, recognize, and infer potential connections in group behavior. To address the issue, inspired by the role of latent spatial position present in video frames, we propose a novel method for learning graph structures by incorporating the distances between individuals. Specifically, we design a graph reasoning module based on Graph Convolutional Networks (GCNs) to learn the hierarchical relationship between individual behaviors and group intentions. To evaluate the feasibility and effectiveness of our proposed model, we conduct experiments on publicly available datasets. Through the experimental results, we validate the effectiveness of our approach, demonstrating its ability to accurately analyze and interpret group behavior.

A Graph Generation Model for Convolutional Neural Network Architecture based on GCN and GAN

In recent years, Neural Architecture Search (NAS) has garnered widespread attention in the field of deep learning due to its significant potential in automating the construction of deep models. However, existing NAS methods primarily focus on optimizing network architecture, utilizing search strategies to find a high-performing network architecture within the search space as effectively as possible. And this process often requires repetitive and continuous searching and evaluation. With the significant advancements in Artificial Intelligence Generated Content (AIGC), an increasing number of researchers are utilizing deep generative models to create graph data. Neural network architecture can be viewed as Directed Acyclic Graphs(DAG) with labeled nodes. Therefore, we propose a graph generation model based on Graph Convolutional Network (GCN) and Generative Adversarial Network(GAN) to generate network architecture. With the aim of avoiding the repetitive and continuous searching and evaluation process in NAS. The CNN architecture generated by our algorithm in this paper achieves an accuracy of 94.37% on the CIFAR-10 dataset. While it may not outperform many other CNN models in terms of performance, it doesn’t require any expert knowledge and is generated automatically by the model, avoiding the need for repetitive searching and evaluation.

Lightweight and efficient skeleton-based sports activity recognition with ASTM-Net

Human Activity Recognition (HAR) plays a pivotal role in video understanding, with applications ranging from surveillance to virtual reality. Skeletal data has emerged as a robust modality for HAR, overcoming challenges such as noisy backgrounds and lighting variations. However, current Graph Convolutional Network (GCNN)–based methods for skeletal activity recognition face two key limitations: (1) they fail to capture dynamic changes in node affinities induced by movements, and (2) they overlook the interplay between spatial and temporal information critical for recognizing complex actions. To address these challenges, we propose ASTM‑Net, an Activity‑aware SpatioTemporal Multi‑branch graph convolutional network comprising two novel modules. First, the Activity‑aware Spatial Graph convolution Module (ASGM) dynamically models Activity‑Aware Adjacency Graphs (3A‑Graphs) by fusing a manually initialized physical graph, a learnable graph optimized end‑to‑end, and a dynamically inferred, activity‑related graph—thereby capturing evolving spatial affinities. Second, we introduce the Temporal Multi‑branch Graph convolution Module (TMGM), which employs parallel branches of channel‑reduction, dilated temporal convolutions with varied dilation rates, pooling, and pointwise convolutions to effectively model both fine‑grained and long‑range temporal dependencies. This multi‑branch design not only addresses diverse action speeds and durations but also maintains parameter efficiency. By integrating ASGM and TMGM, ASTM‑Net jointly captures spatial–temporal mutualities with significantly reduced computational cost. Extensive experiments on NTU‑RGB + D, NTU‑RGB + D 120, and Toyota Smarthome demonstrate ASTM‑Net’s superiority: it outperforms DualHead‑Net‑ALLs by 0.31% on NTU‑RGB + D X‑Sub and surpasses SkateFormer by 2.22% on Toyota Smarthome Cross‑Subject; it reduces parameters by 51.9% and FLOPs by 49.7% compared to MST‑GCNN‑ALLs while improving accuracy by 0.82%; and under 30% random node occlusion, it achieves 86.94% accuracy—3.49% higher than CBAM‑STGCN.

Fusion of Improved Retinex Enhancement and Graph Neural Networks for Retinal Image Quality Classification

The original retinal images are often affected by uneven illumination, noise and other factors, resulting in poor classification accuracy. In this paper, the original retinal image is firstly enhanced by the improved Retinex algorithm, and then the global feature extraction is performed on the enhanced image by using the ResNet-12 model to obtain the global feature vector as the node features, and then the local features of different scales are dynamically adjusted by the multi-scale adaptive aggregation module to highlight the effective information of the nodes. A graph convolutional network is utilized to update the node features, and finally the aggregated node features are input to the fully connected level for classification prediction. Experimental results on two public datasets show that the offered model improves the classification accuracy by 3.12%–15.98%, and is able to more accurately classify retinal images of different quality levels.

STHFD: Spatial–Temporal Hypergraph-Based Model for Aero-Engine Bearing Fault Diagnosis

Accurate fault diagnosis in aerospace transmission systems is essential for ensuring equipment reliability and operational safety, especially for aero-engine bearings. However, current approaches relying on Convolutional Neural Networks (CNNs) for Euclidean data and Graph Convolutional Networks (GCNs) for non-Euclidean structures struggle to simultaneously capture heterogeneous data properties and complex spatio-temporal dependencies. To address these limitations, we propose a novel Spatial–Temporal Hypergraph Fault Diagnosis framework (STHFD). Unlike conventional graphs that model pairwise relations, STHFD employs hypergraphs to represent high-order spatial–temporal correlations more effectively. Specifically, it constructs distinct spatial and temporal hyperedges to capture multi-scale relationships among fault signals. A type-aware hypergraph learning strategy is then applied to encode these correlations into discriminative embeddings. Extensive experiments on aerospace fault datasets demonstrate that STHFD achieves superior classification performance compared to state-of-the-art diagnostic models, highlighting its potential for enhancing intelligent fault detection in complex aerospace systems.

A Method for Multimodal Remote Sensing Image Classification

In remote sensing, images are widely used in applications, such as land cover classification, urban monitoring, and disaster management, providing rich information about the Earth's surface. However, due to data heterogeneity and scarcity, different modalities of remote-sensing images often face challenges in classification tasks. The proposed deep learning model for remote-sensing image classification addresses these challenges through multimodal fusion. By combining a convolutional neural network, a generative adversarial network, and a graph convolutional network, the model is organized into three main components: data preprocessing and feature extraction, multimodal data generation and enhancement, and multimodal feature fusion and classification. Experimental results on the Hyperspectral-Light Detection and Ranging Houston2013 dataset and the Hyperspectral-Synthetic Aperture Radar Berlin dataset show that the proposed method significantly outperforms traditional methods and other deep learning models in classification performance, with better stability and robustness.

Spatio-temporal transformer and graph convolutional networks based traffic flow prediction

Traffic flow prediction is a core component of intelligent transportation systems, providing accurate decision support for traffic management and urban planning. Traffic flow data exhibits highly complex spatiotemporal characteristics due to the intricate spatial correlations between nodes and the significant temporal dependencies across different time intervals. Despite substantial progress in this field, several challenges still remain. Firstly, most current methods rely on Graph Convolutional Networks (GCNs) to extract spatial correlations, typically using predefined adjacency matrices. However, these matrices are inadequate for dynamically capturing the complex and evolving spatial correlations within traffic networks. Secondly, traditional prediction methods predominantly focus on short-term forecasting, which is insufficient for long-term prediction needs. Additionally, many approaches fail to fully consider the local trend information in traffic flow data which reflects short-term temporal variations. To address these issues, a novel deep learning-based traffic flow prediction model, TDMGCN, is proposed. It integrates the Transformer and a multi-graph GCN to tackle the limitations of long-term prediction and the challenges of using the predefined adjacency matrices for spatial correlation extraction. Specifically, in the temporal dimension, a convolution-based multi-head self-attention module is designed. It can not only capture long-term temporal dependencies but also extract local trend information. In the spatial dimension, the model incorporates a spatial embedding module and a multi-graph convolutional module. The former is designed to learn traffic characteristics of different nodes, and the latter is used to extract spatial correlations effectively from multiple graphs. Additionally, the model integrates the periodic features of traffic flow data to further enhance prediction accuracy. Experimental results on five real-world traffic datasets demonstrate that TDMGCN outperforms the current most advanced baseline models.

Integrating Graph Convolution and Attention Mechanism for Kinase Inhibition Prediction.

Kinase is an enzyme responsible for cell signaling and other complex processes. Mutations or changes in kinase can cause cancer and other diseases in humans, including leukemia, neuroblastomas, glioblastomas, and more. Considering these concerns, inhibiting overexpressed or dysregulated kinases through small drug molecules is very important. In the past, many machine learning and deep learning approaches have been used to inhibit unregulated kinase enzymes. In this work, we employ a Graph Neural Network (GNN) to predict the inhibition activities of kinases. A separate Graph Convolution Network (GCN) and combined Graph Convolution and Graph Attention Network (GCN_GAT) are developed and trained on two large datasets (Kinase Datasets 1 and 2) consisting of small drug molecules against the targeted kinase using 10-fold cross-validation. Furthermore, a wide range of molecules are used as independent datasets on which the performance of the models is evaluated. On both independent kinase datasets, our model combining GCN and GAT provides the best evaluation and outperforms previous models in terms of accuracy, Matthews Correlation Coefficient (MCC), sensitivity, specificity, and precision. On the independent Kinase Dataset 1, the values of accuracy, MCC, sensitivity, specificity, and precision are 0.96, 0.89, 0.90, 0.98, and 0.91, respectively. Similarly, the performance of our model combining GCN and GAT on the independent Kinase Dataset 2 is 0.97, 0.90, 0.91, 0.99, and 0.92 in terms of accuracy, MCC, sensitivity, specificity, and precision, respectively.

A KeyBERT-Enhanced Pipeline for Electronic Information Curriculum Knowledge Graphs: Design, Evaluation, and Ontology Alignment

This paper proposes a KeyBERT-based method for constructing a knowledge graph of the electronic information curriculum system, aiming to enhance the structured representation and relational analysis of educational content. Electronic Information Engineering curricula encompass diverse and rapidly evolving topics; however, existing knowledge graphs often overlook multi-word concepts and more nuanced semantic relationships. To address this gap, this paper presents a KeyBERT-enhanced method for constructing a knowledge graph of the electronic information curriculum system. Utilizing teaching plans, syllabi, and approximately 500,000 words of course materials from 17 courses, we first extracted 500 knowledge points via the Term Frequency–Inverse Document Frequency (TF-IDF) algorithm to build a baseline course–knowledge matrix and visualize the preliminary graph using Graph Convolutional Networks (GCN) and Neo4j. We then applied KeyBERT to extract about 1000 knowledge points—approximately 65% of extracted terms were multi-word phrases—and augment the graph with co-occurrence and semantic-similarity edges. Comparative experiments demonstrate a ~20% increase in non-zero matrix coverage and a ~40% boost in edge count (from 5100 to 7100), significantly enhancing graph connectivity. Moreover, we performed sensitivity analysis on extraction thresholds (co-occurrence ≥ 5, similarity ≥ 0.7), revealing that (5, 0.7) maximizes the F1-score at 0.83. Hyperparameter ablation over n-gram ranges [(1,1),(1,2),(1,3)] and top_n [5, 10, 15] identifies (1,3) + top_n = 10 as optimal (Precision = 0.86, Recall = 0.81, F1 = 0.83). Finally, GCN downstream tests show that, despite higher sparsity (KeyBERT 64% vs. TF-IDF 40%), KeyBERT features achieve Accuracy = 0.78 and F1 = 0.75, outperforming TF-IDF’s 0.66/0.69. This approach offers a novel, rigorously evaluated solution for optimizing the electronic information curriculum system and can be extended through terminology standardization or larger data integration.

ViT-GCN: A Novel Hybrid Model for Accurate Pneumonia Diagnosis from X-ray Images.

This study aims to enhance the accuracy of pneumonia diagnosis from X-ray images by developing a model that integrates Vision Transformer (ViT) and Graph Convolutional Networks (GCN) for improved feature extraction and diagnostic performance. The ViT-GCN model was designed to leverage the strengths of both ViT, which captures global image information by dividing the image into fixed-size patches and processing them in sequence, and GCN, which captures node features and relationships through message passing and aggregation in graph data. A composite loss function combining multivariate cross-entropy, focal loss, and GHM loss was introduced to address dataset imbalance and improve training efficiency on small datasets. The ViT-GCN model demonstrated superior performance, achieving an accuracy of 91.43\% on the COVID-19 chest X-ray database, surpassing existing models in diagnostic accuracy for pneumonia. The study highlights the effectiveness of combining ViT and GCN architectures in medical image diagnosis, particularly in addressing challenges related to small datasets. This approach can lead to more accurate and efficient pneumonia diagnoses, especially in resource-constrained settings where small datasets are common.

Corrosion resistance prediction of high-entropy alloys: framework and knowledge graph-driven method integrating composition, processing, and crystal structure

The prediction of corrosion resistance in High-entropy alloys (HEAs) faces challenges due to previous machine learning methods not fully capturing the interdependencies between composition, processing, and crystal structure. This study proposes the Composition and Processing-Driven Two-Stage Corrosion Prediction Framework with Structural Prediction (CPSP Framework), which first predicts crystal structure and then combines composition and processing data for corrosion current prediction. A deep learning model, Mat-NRKG, is developed based on the CPSP framework, efficiently integrating composition, processing, and crystal structure data through a knowledge graph and graph convolutional network. Evaluations using the HEA-CRD dataset show that the CPSP Framework outperforms the Composition-Only Prediction Framework (CP Framework) and the Composition and Processing-Based Prediction Framework (CPP Framework). The Mat-NRKG model demonstrates the best performance on the HEA-CRD dataset. Its generalization capability is validated through experiments on five laboratory-synthesized HEAs, highlighting the effectiveness of incorporating prior knowledge into model design for performance prediction.